Return to Answer

For example, with Miller,

$ mlr --csv nest --ivar ';' -f colour file.csv
item,type,colour
item1,itemtype,blue;grey
item2,itemtype,blue
item2,itemtype2,orange;blue
item3,itemtype,blue

For example, the given data could be transformed into CSV:

item,type,colour
item1,itemtype,blue
item1,itemtype,grey
item2,itemtype,blue
item2,itemtype2,orange
item2,itemtype2,blue
item3,itemtype,blue

... or into an equivalent database table, and then worked on from there.

If jq becomes difficult to run for larger inputs (a linear extrapolation suggest it should take me about 54 hours to run a 1 GB data set), you may want to look into processing the data in some way (potentially in a database or at least work with the data in some other format which is not JSON).